DSS Agentic Workflow Automation Analysis¶

This analysis breaks down the DSS agentic development workflow by step type, showing what is automated deterministically, what requires LLM inference, and what requires human involvement.

Step Type Breakdown¶

Step Type	Count	Percentage	Description
Deterministic	7	43.8%	Predictable, repeatable automated steps
LLM	2	12.5%	Non-deterministic AI inference
Human	3	18.8%	Human decision-making required
External	2	12.5%	External service invocation
Tool	1	6.2%	Tool orchestration (MCP server)

Total Entities: 16

Deterministic Steps (43.8%)¶

These steps produce predictable, repeatable outputs given the same inputs:

Entity	Description	Idempotent
`lint_spec`	Validates spec completeness	✅ Yes
`prompt_generator`	Generates LLM context from spec	✅ Yes
`code_validator`	Validates code against spec	✅ Yes
`auto_fixer`	Auto-fixes token violations	❌ No (modifies files)
`compliance_reporter`	Generates compliance report	✅ Yes
`release_gate`	Go/no-go decision	✅ Yes
`certificate_generator`	Generates compliance certificate	✅ Yes

Key Insight: 43.8% of the workflow is fully deterministic and can run unattended.

LLM Steps (12.5%)¶

These steps use AI inference and are non-deterministic:

Entity	Description	Model	Policy
`coding_agent`	Generates component code	claude-opus-4	quality-gated
`semantic_fixer`	Fixes anti-patterns via refactoring	claude-opus-4	quality-gated

Key Insight: Only 12.5% of steps require LLM inference. These are the most expensive and unpredictable steps.

Human Steps (18.8%)¶

These steps require human decision-making:

Entity	Description	Timeout
`human_designer`	Creates design spec, defines principles	48h
`human_reviewer`	Reviews semantic fixes and unfixable issues	4h
`human_approver`	Final release approval	24h

Key Insight: Human involvement is required at 3 points: 1. Design time - Creating the spec (policy decision) 2. Fix review - Approving LLM-suggested fixes (quality gate) 3. Release approval - Final sign-off (accountability)

External Steps (12.5%)¶

These steps invoke external services:

Entity	Description	Timeout	Gap Status
`external_validator_a11y`	WCAG accessibility validation	5m	Available via agent-a11y
`visual_validator`	Visual regression testing	10m	GAP: Not integrated

Key Insight: External validators extend DSS capabilities without adding complexity to the core system.

Automation Maturity Calculation¶

Based on the PIDL spec, we can calculate automation maturity:

Deterministic steps:  7 × 1.0 = 7.0
External steps:       2 × 0.8 = 1.6  (automated but external dependency)
Tool steps:           1 × 0.9 = 0.9  (orchestration)
LLM steps:            2 × 0.6 = 1.2  (non-deterministic but automated)
Human steps:          3 × 0.0 = 0.0  (not automated)

Total weighted:       10.7 / 15 = 71.3% automated

Current Automation Score: 71.3%

With the visual_validator gap fixed, this would increase to ~77%.

Flow Analysis¶

By Phase¶

Phase	Flows	Automation Level
1. Spec Validation	3	Deterministic (human creates spec)
2. Code Generation	3	LLM-driven (non-deterministic)
3. Validation	6	Deterministic + External
4. Fix Loop	6	Mixed (deterministic + LLM)
5. Human Review	3	Human-in-the-loop
6. Release Decision	6	Deterministic (human approves)

Total Flows: 27 (including alternatives)

Critical Human Touchpoints¶

Touchpoint	Why Human Required	Path to Automation
Design spec creation	Creative/policy decision	Always human
Semantic fix review	Quality judgment	LLM confidence thresholds
Release approval	Accountability	Policy decision (could be auto-approved above threshold)

Recommendations to Increase Automation¶

High Impact¶

Visual Validation Integration (+6% automation)
Integrate Chromatic/Percy for visual regression
Current status: L0 (gap noted in spec)
Confidence-Based Auto-Approval (+5% automation)
Auto-approve semantic fixes above 95% confidence
Human review only for low-confidence changes
Release Auto-Approval for Non-Breaking Changes (+3% automation)
Auto-approve releases with 100% compliance score
Human approval only for exceptions

Medium Impact¶

Expand Deterministic Fixes
Add more rule-based auto-fixes beyond tokens
Reduce reliance on LLM for common patterns
External Validator Orchestration
Auto-invoke configured validators
Aggregate results into compliance report

Summary¶

┌─────────────────────────────────────────────────────────────────┐
│             DSS WORKFLOW AUTOMATION BREAKDOWN                    │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  Deterministic    [████████████████████░░░░░░░░░░░░░░] 43.8%    │
│  LLM              [█████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░] 12.5%    │
│  Human            [███████░░░░░░░░░░░░░░░░░░░░░░░░░░░] 18.8%    │
│  External         [█████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░] 12.5%    │
│  Tool             [███░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░]  6.2%    │
│                                                                  │
│  ─────────────────────────────────────────────────────────────  │
│                                                                  │
│  Total Automation Score: 71.3%                                   │
│  Agentic Maturity Level: L2.8 / L5                              │
│                                                                  │
│  Key Gaps:                                                       │
│  • Visual regression testing (L0)                                │
│  • External validator auto-invocation (L1)                       │
│  • Semantic fix auto-approval (L1)                               │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

PIDL Artifacts¶

Spec: dss_agentic_workflow.json - Full PIDL process specification
Diagram: dss_agentic_workflow.mmd - Mermaid sequence diagram

Generate diagrams with:

pidl generate -f mermaid dss_agentic_workflow.json
pidl generate -f svg dss_agentic_workflow.json -o diagram.svg
pidl generate -f infographic dss_agentic_workflow.json -o infographic.svg